3D SENSING APPROACHES FOR PRECISION AGRICULTURE: APPLICATIONS TO PLANT FEATURES CHARACTERIZATION AND TO WEED CONTROL IN MAIZE

This PhD thesis investigates the application of 3D data acquired by Time-of-Flight (ToF) sensors and stereo cameras for agronomic measurements. The main objectives are the characterization of plant geometric properties, biomass estimation with non-destructive methods and discrimination between crops and weeds for use in precision agriculture applications. Following an introductory section, the first part of the work, which corresponds to Chapters 3 and 4, involves the volumetric characterization of lettuce plants to estimate biomass using non-destructive measurements. This was performed both in preliminary laboratory trials and in a field application. In the field application the depth sensor was integrated with a robotic platform to ensure the possibility to collect data at the individual plant level, with all data that can be stored in a data lake. In both cases, the assumed volumetric reconstruction algorithms achieved very satisfactory results, obtaining accurate dry biomass weight (DM) estimates from the developed models for both laboratory scale (with a root mean square error (RMSE) of 0.12 g and a mean absolute percentage error (MAPE) of 12%.) and greenhouse conditions (with a RMSE of 1.07 g and MAPE 16%). The different ability to estimate the DM obtained in the two contexts, although minimal, depends mainly on the different spatial resolutions obtainable with the acquisition setups designed for the two experiments and on the conditions under which the measurements were carried out. Considering the technological characteristics of the ToF sensors, under field conditions they are disturbed by solar radiation, which contributes to decreasing the level of accuracy of the measurement obtained. The second part of the thesis, which corresponds to Chapters 5, 6, 7 and 8, deals with the comparison of two sensing techniques, two-dimensional based on color (RGB) and three-dimensional that consider distance in addition to color (RGBD) to optimize the process of target identification for use in precision weed control spraying that aims to reduce chemical inputs using advanced techniques. Regarding 2D imagery, a comparative analysis was conducted on images captured using drones (UAV – Remote sensing) and those obtained in proximity to vegetation (proximal sensing). The objective was to assess the capacity of these approaches to estimate soil cover, a critical driver that can indirectly indicate the degree of weed infestation. Proximal images demonstrated superior accuracy in estimating this phenomenon, achieving a coefficient of determination (R²) of 0.93, compared to UAV-acquired images, which yielded an R² of 0.8 in validation tests based on 90 ground control points established within the field. When comparing the two monitoring techniques at a field scale, UAV imagery was found to provide an estimate of soil cover that, in 75% of the sampled area, was lower than that derived from proximal images under experimental conditions. In this context, which highlights the superior representativeness of information obtainable through proximal monitoring, the contribution of 3D sensing techniques to estimating weed presence was subsequently evaluated in comparison to proximal 2D images. To this end, several algorithms specifically designed to utilize both sensing techniques were developed. Using 2D images, two algorithms were developed. The first is a simpler approach that enables uniform treatment decisions across areas of 0.75 m² (Management Zones - MZ) based on assessing vegetation soil cover. The second, an innovative algorithm known as Local Cover Fraction (LCF), is more complex, employing a hybrid technique to simultaneously detect weeds between rows and subsequently characterize the spatial distribution of remaining soil cover along each crop row. This allows identification of areas where biomass distribution deviates from the clean-field conditions recorded at the experiment's start. In parallel, a classification algorithm based on 3D sensing was developed, with two distinct implementation strategies proposed. The first, Treat All Weeds (TAW), aims to target all areas identified as weeds and the second, Treat Weed in Majority (TWM), activates actuators only when they can cover a greater surface area classified as weeds than of crops. To evaluate the results and quantify the contribution of these methods compared to traditional uniformly distributed (UD) techniques, precision spraying treatments based on the MZ, LCF, and 3D-based algorithms were simulated, and the outcomes were compared. Simulations yielded promising results, forecasting a reduction in both the quantity of herbicide applied and the environmental impact for most developed algorithms. Herbicide reductions were observed as follows: 10% for the MZ method, 45% for the LCF method, and for the 3D sensing approach, a 22% reduction for the TAW strategy and 28% for TWM. However, the TAW strategy was the only approach to achieve a statistically significant reduction in both herbicide quantity and environmental impacts without a notable decrease in treatment effectiveness, which remained statistically comparable to the uniformly distributed (UD3D) treatment simulation. The results obtained with both imaging techniques surpassed expectations. Compared to the 2D imaging-based method for estimating weed control treatments, the 3D survey produced even more favorable results, owing to its enhanced capability for target classification, which proved more robust than the 2D method. This advanced classification capability enabled the design of crop-targeted treatments through simulations of localized biostimulant distribution, achieving full crop surface coverage with a sensitivity of 0.99. This approach reduced product waste from ground distribution by 85% and minimized the undesired effect of weed biostimulation by 75%. Finally, in Chapter 9, a low-cost simulator prototype is proposed for the hardware and software management of highly precise, localized treatments based on proximal imaging. This simulator is designed to perform real-time monitoring and execute the required treatment actions. In general, considering the results obtained in both fields of application investigated in this thesis work, the use of 3D sensors demonstrated significant outcomes across all evaluated contexts, both in initial laboratory tests and in field application trials.